AI is Eating Software

‍

SaaS was simple: build something once. Sell it an infinite number of times at de minimis marginal cost. Price it at 80% gross margins. Lock customers in for high retention. Grow exponentially. Enjoy annuity streams forever. 

‍

This is, frankly, the social contract of the past ~15 years of venture capital. Without massive capital leverage, it’s not possible to generate venture-style returns. 

‍

For me, the ultimate business model is the biggest open question in AI. AI (as it is today) breaks the fundamental contract of the software business model. Now the marginal cost isn’t zero - it’s proportional to usage. At scale, will these businesses look more like Spotify, DVDs-in-the-mail Netflix, or Planet Fitness, where your best users are your worst customers (until we get really good at pricing)?

‍

At the application layer we’re seeing 3 business models:

‍

Insane high-growth subscription products

Think Cursor, Lovable, and many others. Growing at never-before-seen-rates. Genuinely astounding.  But they are not low marginal cost - every time a customer uses the product, they also use tokens. More usage equals more tokens. The customers who love and use you the most end up costing you the most money, like the subscribers who rented the most DVDs. 

‍

Yes, tokens get cheaper over time, but human nature drives us to the best, newest tokens, which command a high price while the older prices go down. Of course, pricing will evolve, simple one-size-fits-all subscriptions won’t last forever, infrastructure will change, and moats can form. But today, the fastest growers here are generally single-player or bottoms up tools, which brings retention into question. Easy come, easy go.

‍

‍

I’ll be honest. This is where I’m having the hardest time as a VC, and it’s what’s sucking the air out of the room. 

‍

But what am I missing? Very smart people love these companies. Is the bull case that winner-take-all dynamics drive pricing power that makes margins irrelevant? Or that economics shift when open source models run at the edge? Or that these companies are collecting very valuable datasets? 

‍

Maybe my real question is: am I skeptical of the model, or the deals? Different concerns, but both impact outcome. 

‍

Software with AI inside

To get software margins while being AI-native, you have to put something between the front-end and the inference: coordination, workflow, context, multi-player collaboration, etc. 

‍

These companies are top down enterprise sales by necessity, and it’s physically impossible for them to grow as fast as a single player tool. Humans move slower than technology. Committees of humans are comparatively glacial. Since these companies can’t put up the same numbers as the group 1 rocketships, they’re not getting the same VC attention. 

‍

We call these “AI Onramps”, companies that make AI practical and accessible for specific business needs by embedding into existing workflows, orchestrating human-AI handoffs in high-stakes environments, and selling measurable outcomes rather than AI access or broad efficiency gains. We have invested in several. The thesis is that most enterprises won’t adopt AI through a single magical tool. They will adopt it through software that fits how they already work and solves problems they already have. But these onramps face pressure from two directions: incumbents pushing AI onto their install bases, and enterprises wondering if/when they can just build it themselves. 

‍

I think there’s still time to build moats before robots build our software from prompts we thought of in passing. Coordination, standardization, and organizational buy-in will continue to be hard even as creating software gets easier. I would be lying to you if I knew how it will play out when agents are the ones selecting tools and software. (A bigger question is who inside the org will actually select and guide the agents.) But I think the opportunity today is to go after the workflows that are too messy and cross functional for DIY and become a data- and context-rich connective tissue across systems, reflected in our investments in companies like ASI & Freeplay, among others. 

‍

The open question is how long the window will be for this opportunity. If incumbents move faster or acquire more than expected, how do you position yourself not to get squeezed? Is “AI Onramp” a durable category or a transitional one? How do these companies build something lasting and not become bridges that get washed out when enterprises figure out AI and agents for themselves?

‍

AI-enabled services

These companies sell outcomes, not software (generally to a buyer who doesn’t usually buy software) - think of an AI structural engineering firm or an AI fund administrator. To the customer, they look like a new kind of professional services firm and compete with consultants, agencies, or outsourced labor, rather than SaaS vendors.

‍

There are real benefits here: these businesses aren’t trying to compete with incumbents stuffing mediocre AI into their channel. Their services competitors can’t build internal software (yet). And the output is generally something that the customer absolutely needs and already spends real money. A stamped drawing in hours instead of weeks. An X-ray read instantly. This is where AI will viscerally feel like magic. 

‍

The bet here is that models do more of the work - better, faster, and cheaper - over time, driving down marginal cost and creating software-style economics.The risk is that they never quite get there, that AI behaves more like specialized labor than software, and productivity gains get competed away faster than leverage is captured. 

‍

I’m personally very excited about this category, and we have invested in a couple of these. By being vertically integrated and compounding on their learning, these companies have the opportunity to reach new levels of quality at never-before-seen scale. However, I do wonder how do these companies position themselves to capture the productivity gains they create and not give them all to their customers or the market?

‍

Now what?

‍

I’m still working through how this evolves, but I strongly believe that the returns will be there. They’re just not going to look like SaaS returns. The social contract that powered the last fifteen years of venture is broken, and the opportunity now is to figure out what’s going to replace it. I’m watching Group 1 for pricing and cost structures that change the margin math. Group 2 for what startups can do that incumbents and customers cannot. Group 3 for where the productivity gains accrue. More to come.

‍